REVEALING MALFEASANCE: HOW LOCAL MEDIA

FACILITATES ELECTORAL SANCTIONING

OF MAYORS IN MEXICO ∗

HORACIO A. LARREGUY † JOHN MARSHALL ‡

JAMES M. SNYDER JR. §

AUGUST 2014

We estimate the effect of local media outlets on political accountability in Mexico,focusing on malfeasance by municipal mayors. We study federal grants earmarkedfor infrastructure projects targeting the poor, and leverage two sources of exogenousvariation. First, we exploit variation in the timing of the release of municipal au-dit reports. Second, going beyond existing studies, we exploit within-municipalityvariation in media at the electoral precinct level. In particular, we use a geographicregression discontinuity design, comparing neighboring precincts on the boundariesof media stations’ coverage areas, to isolate the effects of the media environment onvoter behavior. We find that voters punish the party of malfeasant mayors, but onlyin electoral precincts covered by local media stations. Each additional local radio ortelevision station reduces the vote share of an incumbent political party revealed to becorrupt by 1 percentage point, and reduces the vote share of an incumbent politicalparty revealed to be have diverted funds away from the poor by around 2 percentagepoints. The electoral costs of diverting resources away from the poor are especiallylarge for the populist PRI party. However, we find no effect of media stations thatcover the municipality but are based in other municipalities.

JEL: D72, D78, H41, O17.Key words: elections, voter behavior, corruption, malfeasance, media.

∗Thank to Andrea Ortiz and Daniel Silberwasser for excellent research assistance, and to ASF officials for pro-viding information about the auditing process. Horacio Larreguy acknowledges financial support from the IQSSUndergraduate Research Scholars Program. All errors are our own.†Department of Government, Harvard University, Cambridge, MA (hlarreguy@fas.harvard.edu).‡Department of Government, Harvard University, Cambridge, MA (jlmarsh@fas.harvard.edu).§Department of Government, Harvard University, Cambridge, MA (jsnyder@fas.harvard.edu).

1

1 Introduction

A large body of scholarship in political economy asserts that in large democracies: (i) electionsare one of the key institutions for producing political accountability; (ii) in order for electionsto function well, voters must be adequately informed; and (iii) the mass media play an essentialrole in informing voters.1 One important application of this trio, which is especially important indeveloping democracies, is the electoral sanctioning of corruption and other malfeasant behavior.

There is, however, relatively little solid evidence that the media actually inform significantnumbers of voters about the behavior of malfeasant politicians. The strongest evidence is fromFerraz and Finan (2008), who find that incumbent mayors in Brazil who are revealed to be corruptsuffer more at the polls in areas with local radio stations than in other areas. Other studies find thatcorrupt politicians are more likely to be punished electorally when their corruption is covered inthe news, or when political corruption is more generally salient.2 However, none of these studiesexploit exogenous variation in media or the salience of the media coverage.

In addition, it is not clear whether voters care about malfeasance, and if so, what types ofmalfeasance matter most to them. Some studies find that when incumbent politicians are exposedas corrupt then either the incumbents themselves or the parties to which they belong receive sig-nificantly fewer votes at the next election.3 However, other studies find mixed results, or no signif-icant relationship between evidence of corruption or malfeasance and vote shares.4 Furthermore,although corruption may not affect voter behavior, other aspects of politicians’ performance—suchas overall rankings, or targeting public spending toward the poor—does.5 Theorists have respondedby providing arguments to explain why voters might not punish corrupt politicians.6

In this article we estimate the effect of local media outlets on political accountability in Mexico,focusing specifically on municipal mayors. Our detailed local data allow us to exploit two sources

1See, e.g., Thomas Jefferson: “The functionaries of every government have propensities tocommand at will the liberty and property of their constituents. There is no safe deposit for thesebut with the people themselves, nor can they be safe with them without information. Where thepress is free, and every man able to read, all is safe” (Thomas Jefferson to Charles Yancey, 1816.ME 14:384).

2See, e.g., Chang, Golden and Hill (2010), Costas, Sole-Olle and Sorribas-Navarro (2011), andEggers and Fisher (2011) for studies of Italy, Spain and England respectively.

3See, e.g., Ferraz and Finan (2008), Slomczynski and Shabad (2012), and Banerjee et al. (2014).4See, e.g., McCann and Domınguez (1998), Dutta and Gupta (2012), de Figueiredo, Hidalgo

and Kasahara (2013), and Chong et al. (forthcoming).5See, e.g., Banerjee et al. (2011) and Humphreys and Weinstein (2012).6See, e.g., Rundquist, Strom and Peters (1977), Caselli and Morelli (2004), Besley and Prat

(2006), Dutta and Gupta (2012), and Svolik (2013).

2

of plausibly exogenous variation. First, as in Ferraz and Finan (2008), we exploit variation in thetiming of the release of municipal audit reports. In particular, we compare mayors who engagein malfeasant behavior that is revealed in audit reports published in the year before an election tosimilar mayors whose audit reports are not published until after the election. Second, and goingbeyond existing studies, we leverage within-municipality variation in the electoral precincts thatare covered by radio and television stations located within the municipality. To ensure that this isnot correlated with precinct-level differences in development, we use a geographic regression dis-continuity design to compare neighboring precincts at the boundaries of stations coverage areas.7

Furthermore, we extend existing studies by distinguishing between different types of malfeasantbehavior.

A significant proportion of government spending is administered by Mexico’s c.2,400 mayors.Given widespread concerns about corruption, in 1999 the Mexican Congress passed a law institu-tionalizing independent audits of the use of federal funds. Importantly, such audits are announcedafter the federal funds have been disbursed. We focus on the audit reports pertaining to the Munic-ipal Fund for Social Infrastructure (FISM). FISM is a major social program that provides mayorswith funds for infrastructural projects intended to benefit impoverished citizens and represent about25% of mayors’ annual budgets. The audit reports reveal the share of FISM money spent “in anunauthorized manner,” as well as the share spent on projects “not benefiting the poor.” The firstfigure clearly represents malfeasance, and usually actual corruption. The second figure indicatesmalfeasance of a different sort—diverting funds from their intended targets. By law, 100% ofFISM projects must benefit the poor, so any money not spent on the poor represents misallocation.

Our results demonstrate that voters punish the party of malfeasant mayors, but only in electoralprecincts covered by local media stations. The point estimates imply that each additional localradio or television station reduces the vote share of an incumbent political party revealed to becorrupt by one percentage point.8 The effects of misallocating funds away from the poor are evenlarger. Our point estimates imply that if the incumbent party has significantly misallocated fundsaway from the poor, each additional local radio or television station reduces the party’s vote shareby between 1.5 and 2.5 percentage points, depending on the severity of the malfeasance. However,these effects only hold for media stations based within the municipality to which a given auditreport applies. We find no effect of media stations that cover the municipality but are based inother municipalities.

We also find evidence suggesting that voters punish parties more for behavior that is not only

7These are defined by the areas reached with commercial quality signal strength.8We examine the electoral implications of malfeasance for the incumbent mayor’s political

party, because incumbent mayors could not run for re-election themselves due to term limits.

3

malfeasant but contrary to the party’s ideological reputation. More specifically, voters punishthe PRI for diverting FISM funds away from the poor, but do not punish the PAN. The PRI hasa stronger “pro-poor” reputation than the PAN.9 Thus, it might be more surprising when a PRImayor is caught misallocating FISM funds by spending on the non-poor, making it more likely thatvoters significantly update their beliefs about the party’s sincerity, commitment, or competence.Alternatively, voters might simply find this behavior by the PRI to be particularly hypocritical,therefore more egregious and deserving of punishment.

We show that our results are robust to a series of sensitivity checks. First, our results are robustto using an alternative measures of corruption and misallocation of funds away from the poor,as well as alternative definition of party incumbents, which might be of relevance since parties inMexico often run for office on short-lived coalitions. Second, through the inclusion of municipalityfixed effects or by simply focusing on within-municipality matches, we show that our estimates arenot driven by matches where the treated and control units come from neighboring municipalities.Lastly, we examine the effect of different types of media separately and find that local media’seffect are largest for FM radio and television.10

Our findings contribute to the literature in a variety of ways. First, we exploit a source ofvariation for estimating media effects that has not been explored in the literature on political ac-countability in developing democracies. A number of previous studies have applied the samegeneral idea—using detailed features of the media market environment to obtain plausibly exoge-nous variation in the degree of “exposure” to different media outlets or messages—to primarilystudy phenomena such as the impact of media bias.11 Snyder and Stromberg (2010) and Fergusson(2014) also apply this idea to study how the media market environment improves accountability inthe U.S., but do not provide evidence of a direct link to political malfeasance.

Some of the studies that focus on the degree to which voters respond to corruption and othertypes of malfeasance also provide evidence of a media linkage, but as noted above, the evidenceis often more suggestive than conclusive. Ferraz and Finan (2008) find that the number of localAM radio stations increases the electoral response to mayoral corruption in Brazil. As the authorsacknowledge, however, the presence or absence of local stations may be correlated with a variety

9For example, according to the 2009 CIDE-CSES survey, voters identify the PRI to the left ofthe PAN. See the Comparative Manifesto Project codings here.

10We also show that our estimates are not driven by differences in internet access—a variablethat our matches present an imbalance on—by controlling for a triple interaction between auditresults and the proportion of the precinct with internet access.

11See, e.g., DellaVigna and Kaplan (2007), Chiang and Knight (2011), and Enikolopov, Petrovaand Zhuravskaya (2011). See DellaVigna and Gentzkow (2010) for a review of this literature.

4

of municipal characteristics, only some of which can be measured and included as controls. Aparticular concern is that more developed municipalities, which are more likely to be have localradio stations, are also more likely to contain better educated and more politically engaged andsophisticated voters, who may be more sensitive to mayoral performance.12 Thus, although Ferrazand Finan (2008) control for a battery of municipal-level variables, their empirical strategy cannotrule out the possibility that their estimates are biased due to the presence of such unobservedconfounders.

Similarly, Eggers and Fisher (2011) find a statistically significant loss in vote share for legis-lators whose misbehavior in the 2009 expenses scandal was serious enough to be featured in thenews, but not for other implicated legislators. Costas, Sole-Olle and Sorribas-Navarro (2011) alsofind that the incumbent’s vote share loss was larger for corruption cases that were widely reportedin the newspapers. While these findings are consistent with a media effect, it is also likely that thecorrelations reflect differences in the types of misbehavior covered by the media. In particular—asthe authors of these studies clearly recognize—media outlets are likely to cover “serious” malfea-sance more heavily than “minor” malfeasance. Chang, Golden and Hill (2010) show that therewas more coverage of political corruption and political party finances in one of the major Italiannational newspapers during the election for which they find a significant negative effect of criminalcharges on voting (the 1994 election), relative to the previous eight years.

Unlike these previous studies, our estimates almost surely understate, rather than overstate, theeffects of providing more media to voters. More precisely, we provide estimates of the “intention-to-treat” voters with access to media, because our geographic regression discontinuity is fuzzyrather than sharp.13 Some of the precincts that we classify as outside the coverage area of a givenstation are actually “partially inside” the station’s coverage area, and that some of the precinctswe classify as inside a station’s coverage are are only “partially outside.” As a result, the averagedifference in reception between the precincts that are just inside and just outside of a station’scoverage area is less than complete even though we treat it as if it were complete.

Second, we demonstrate the importance of local media for local political accountability, rather

12Klasnja (2011) and Weitz-Shapiro and Winters (2014) find evidence that voters with greater“political awareness” or literacy are more likely to punish incumbents in scandals.

13As we discuss in more detail below, radio and television signal strengths decay gradually asa function of geographic distance, not abruptly, and are also affected by terrain, large obstacles,ground conductivity, and even weather and temperature. Whether or not a given household canreceive a signal may also depend on the equipment—e.g. the type of antenna—the householdhas. So, the exact boundaries of each station’s coverage area are never perfectly accurate (or evencompletely fixed). Another issue, especially for radio, is that many people who commute may hearradio stations outside the area where they live.

5

than media in general. This is an important consideration because local radio and television areoften the only way in which isolated voters can learn about the performance of their incumbentpoliticians, and around 20% and 25% of Mexican electoral precincts are respectively not coveredby a single FM or television station.14 Moreover, the importance of local media is of particularimportance since local media markets are shrinking in many countries. In Mexico, in the last15 years there was a 40% decline in the share of individuals claiming to read political news innewspapers.15 In the U.S., daily newspaper circulation dropped from just over 1.0 newspapers perhousehold in 1950 to about 0.3 per household in 2010. This trend is particularly worrying giventhat our evidence provides a clear rationale for politicians to exploit the weakening economicposition of local media and to seek its control (Besley and Prat 2006), by purchasing radio stations(Boas and Hidalgo 2011) or preventing “defamation” (Stanig forthcoming).

Third, we show that voters respond differently to different types of malfeasance. As notedabove, the bulk of the literature on political accountability in developing democracies has focusedon corruption. Two exceptions are Banerjee et al. (2011) and Humphreys and Weinstein (2012).Humphreys and Weinstein (2012) conducted a field experiment in Uganda, and find that provid-ing voters with information about the overall performance of their incumbent legislator relativeto other legislators—as measured by an index involving participation in floor debates and votes,participation in committee debates and votes, and constituency service—leads voters to updatewhether they approve of, or intend to vote for, their incumbent. Banerjee et al. (2011) conducteda field experiment in India, and find that voters living in slums are more likely to vote for theirincumbent legislator if they learn that the incumbent allocates more of her discretionary projectspending funds to slums rather than other areas. Our findings are analogous to those of Banerjeeet al. (2011), since we find that voters are more likely to vote against incumbents who divert fundsaway from the poor.16

Fourth, since audits are announced only after FISM funds have been allocated, our resultsshow that the possibility of being audited is insufficient to prevent municipal mayors from engag-ing in malfeasance. Our findings thus complement previous research suggesting that audits canbe effective at reducing corruption if it is known before spending occurs that the reports couldresult in criminal prosecution (Olken 2007) or be released before an election (Bobonis, Fuertes

14Local radio is less prevalent in Brazil (Ferraz and Finan 2008).1560% of Latinobarometer respondents claimed reading political news in newspapers 1996,

compared to 36% in 2009.16The analogy is not perfect, since in our case the diversion of funds away from the poor is a

direct violation of FISM program rules, while in their case it is not—legislators in India are free toallocate their discretionary project funds anywhere in their districts.

6

and Schwabe 2014). Rather, the corruption levels we find in Mexico are broadly similar to thoseobserved by Ferraz and Finan (2008) in Brazil, where the municipal audit scheme was only an-nounced after spending had occurred. Our results thus suggest that it is only the certainty of beingaudited that causes politicians to anticipate being audited and alter their malfeasant behavior. Thisis consistent with the dynamic optimizing behavior observed in India (Niehaus and Sukhtankar2013).

The article proceeds as follows. Section 2 provides a brief overview of local governments inMexico, the FISM funds that we study, the audit of such funds, and local media in Mexico. Section4 details our data and identification strategy. Section 5 presents our main results and robustnesschecks. Section 6 concludes.

2 Political accountability in Mexico

2.1 Municipality audits

Mexico’s 31 states contain around 2,400 municipalities, which are currently responsible for 20% oftotal government spending after a significant fiscal decentralization that took place in the 1990s.17

Municipal governments are led by mayors, who are responsible for delivering basic public servicesand managing local infrastructure. Mayors are normally elected every three years, although theyserve four-year terms in some states, and could not stand for re-election.18

An important component of a mayor’s budget is the Municipal Fund for Social Infrastructure(FISM). On average, this represents 24% of a municipality’s total income. FISM funds, whichare allocated to municipalities by the Fiscal Coordination Law (LCF) passed in 1997, are directfederal transfers provided exclusively for the funding of public works, basic social actions, andinvestments that directly benefit the socially disadvantaged population that is in extreme poverty.Spending can be assigned in any of the following categories: potable water, sewage, drainageand latrines, municipal urbanization, electrification or rural and poor suburban areas, basic healthinfrastructure, basic education infrastructure, improvement of housing, rural roads and rural pro-ductive infrastructure. Unlike previous studies focusing on corruption in more general programs(e.g. Ferraz and Finan 2008; Bobonis, Fuertes and Schwabe 2013), the specific targeting of FISMfunds allows us to examine the electoral response by voters to both corruption and the misuse offunds intended to serve a disadvantaged population.

17Education and health were decentralized between 1992 and 1996 and the decentralization ofinfrastructure projects followed in 1997 (Wellenstein, Nunez and Andres 2006).

18Re-election will become possible for those running starting in 2015.

7

Compared to previous social programs, FISM funding has been relatively successful at tar-geting resources at the poor (Wellenstein, Nunez and Andres 2006). However, funds are oftenmisallocated. In some cases, mayors engage in corrupt activities such as diverting resources, notoffering public contracts up for open tender, and expensing personal largesse. FISM transfers canalso spent be on public works not targeted at poor areas, such as paving the streets or improvingsewage in rich urban areas.

The use of FISM funds is subject to independent audits by Mexico’s Federal Auditor’s Office(ASF). The ASF, which was established in 1999 in response to widespread concerns regardingthe mismanagement of public resources, is an independent body with constitutionally-enshrinedpowers to audit the federal, state and municipal governments. In each year since 2000, the ASFhas audited FISM spending in multiple municipalities per state. Importantly, an audit is announcedafter the spending has occurred. Although the exact formula for selection is not publicly available,official information indicates that municipalities are chosen on the basis of the federal transfersthey received, the importance of these funds relative to the municipal budget, whether they havebeen audited before, and their history of misallocated expenditure.19 However, our identificationstrategy exploits the timing of audits, rather than comparing audited and non-audited municipali-ties. The ASF’s selection rule defines the population to which our estimates apply.

Audits focus on the spending and management of FISM resources, and are conducted by in-dependent ASF officials. Auditors check that officials abide by the rules established for the man-agement of FISM resources (e.g., procurement rules, accounting procedures), that the status of thefunded projects is in accordance with the books, and that funds are given the use they were intendedfor. Audit reports are not publicly released until up to two years after the audit was conducted. TheASF presents to Congress the results of all audits conducted in the calendar year two years prior bythe last working day of February each year. The reports are then made publicly available online.

ASF audit reports break down the use of FISM funds across several dimensions. Most impor-tantly, the reports state the percentage of FISM funds spent on infrastructure projects not bene-fiting the poor and the percentage of funds used for unauthorized spending. Spending that doesnot benefit the poor ranges from the diversion of resources to support agricultural production (dur-ing election times) to paving the streets of relatively rich urban areas. We interpret unauthorizedspending, which includes the diversion of resources of personal expenses of the mayor and fundsthat are unaccounted for, as corruption.20 In the Online Appendix, we provide an example of an

19Personal interview with the Licentiate Jaime Alvarez Hernandez, General Director of Researchand Evaluation of the Special Audit of Federal Spending, in July 2012.

20This definition resembles Ferraz and Finan (2008) in that we focus on violations that includeprocurement fraud, diversion and over-invoicing, but differs in that we quantify the relative impor-

8

audit report.The ASF can impose a variety of punishments on malfeasant public officials. In particular, the

ASF can inflict municipality fines to recover FISM funds, recommend that the Ministry of PublicFunction removes, suspends or imposes economic sanctions on officials, or file (or recommend)a criminal case against culpable individuals. However, in practice these punishments have notbeen used regularly: between December 2006 and July 2012, the Ministry of Public Function onlyrecovered two million U.S. dollars, sanctioned 9,000 public employees for serious misdemeanors,and incarcerated one hundred officials.21

2.2 Local media

As in many developing countries, radio and television networks are the principal source of newsin Mexico. While a few television stations are predominantly national in focus, the focus of radiostations and most television stations is predominantly local.

The annual release of municipal audit results in February is a much-anticipated media event.This is particularly true for television networks, which widely broadcast footage of the release.However, many news reports concerning the ASF releases are published in February and March ofeach year. Supporting this claim, Figures 1 and 2 show that trends in Google searches for the ASFand FISM spike around February and March each year. Nevertheless, we still found many newsreports from later in the year.

[Figures 1 and 2 about here.]

Media reports, which generally cover mayors within a given state or just the local vicinity,almost exclusively focus on cases of corruption and mayors not spending FISM funds on projectstargeting the poor. Most reports cite the exact proportions of unauthorized spending and spendingon projects not targeted at the poor.22 Little mention was made of other features of the reports suchas the the degree of participation of the community in the allocation of funds or the share of FISMfunds that were spent, which are also other variables disclosed in the ASF reports.

tance of such corruption. Rather than the percentage of unauthorized spending, Ferraz and Finan(2008) count the number of corruption violations.

21El Universal, “A la carcel, solamente 100 ex servidores”, 29th May 2014, link.22For example, see: BBM Noticias, “ASF: desvio Ugartchechea 370.9 mdp”, October 21st 2013,

here; El Informador, “Hallan irregularidades en gasto tapatıo contra pobreza”, February 28th 2013,here; Revolucion Tres Punto Cero, “En 2012, se desviaron a campanas 29 millones de pesos paracombate a la pobreza en Tabasco”, March 6th 2014, here.

9

The bulk of news reports focus on particularly egregious cases of corrupt and neglectful may-ors.23 For example, in 2013 it was revealed Oaxaca de Juarez’s mayor had created a fake union tocollect payments, presided over many public works contract without offering open tender, divertingpayments for advertising and consulting fees, and failed to provide details of considerable quanti-ties of spending.24 While Mayor Luis Julian Ugartechea Begue represents one of the most corruptmayors, such behavior was not uncommon. Many reports pointed to mayors diverting payments,using FISM funds for personal and family expenses and manipulating tender processes. Failures tospend FISM funds on the poor were just as common in media reports. In many cases, public worksprojects were undertaken in urban and affluent parts of the city. In others, the alleged project nevermaterialized despite being paid for, or was diverted for alternative uses such as supporting localcandidates from the incumbent’s party.

Mayors in Mexico cannot stand for re-election. This feature differentiates our study from manypreceding studies (e.g. Banerjee et al. 2011; Ferraz and Finan 2008; Humphreys and Weinstein2012). However, there are good reasons to believe that a mayor’s political party may still bepunished by voters at the next election. First, political parties do not disappear with mayors.Although the local coalitions between parties uniting behind a mayoral candidate can change acrosselections, political parties always back a particular candidate. Second, previous evidence showsthat at least in some cases political parties are punished for the actions of their leaders (e.g. Chonget al. forthcoming).

Voters in Mexico appear to be generally unaware of mayoral responsibilities (Chong et al.forthcoming). Moreover, most public spending is invisible and inaccessible to most voters. A fewvoters will know about a few public projects from direct experience; for example, a voter who visitsa new hospital or health clinic due to an illness, a voter with children who attend a newly renovatedschool, or a voter who drives over a newly paved road (after months of frustration driving aroundthe construction site). For the most part, however, voters only learn about public spending throughmedia coverage and thanks to the ASF audit reports.25 This is even more true of money not spent

23Because of this, media coverage may be biased toward large municipalities where the sumsinvolved are large in nominal terms or smaller municipalities where the proportion of FISM fundsmisallocated is especially large.

24BBM Noticias, “ASF: desvio Ugartchechea 370.9 mdp”, October 21st 2013, here.25In principle, local governments are required to inform the public about the arrival of FISM

funds. However, only about 50% comply with this requirement. Moreover, among those that docomply, the main communication channels used are newspapers and the internet—i.e., two typesof mass media. Furthermore, media relies extensively on the ASF audit reports since governmentsare extremely reluctant to release information about their expenses to the public (Lavielle, Pirkerand Serdan, 2006).

10

that should have been spent, or money that is spent improperly. Survey data is consistent with thisview. According to the 2009 Latinobarometer, for example, 83% of respondents gather politicalinformation from TV, 41% gather political information from radio, 30% gather political informa-tion from newspapers, and 41% gather political information from family, friends and colleagues(many of whom, of course, gather their information from TV, radio and newspapers).26

We now turn to our data, and to the empirical strategy we use to identify the effects of therelease of audit reports in electoral precincts with and without local media outlets.

3 Data

This section describes our main sources of data: electoral results at the electoral precinct level;municipality audit reports; and precinct-level radio and television coverage.

3.1 Mayoral election outcomes

Mexico’s municipalities (and Congressional districts) are divided in to around 67,000 electoralprecincts. Using data from the Federal Electoral Institute (IFE) and State Electoral Institutes, wecollected electoral returns for every available precinct in each municipal election between 2002and 2012. States hold municipal elections in different years, and at three of four year intervals. Wethus accumulated up to four election results per electoral precinct, which enabled us to identify theincumbent and incumbent’s past vote share in all the elections in our period of analysis, 2007-2012.

We focus on two main outcome variables: the change in incumbent party’s vote share at theprecinct level, and whether the incumbent party was re-elected at the municipal level. The formermeasure quantifies the extent of voter sanctioning, and the latter captures the implications for theidentity of the office-holder. We define the vote share as a proportion of voters that turned out.27

Mexico’s mayors could not stand for re-election, so we focus on the party of the incumbentmayor. Given municipal politics entails the formation of local coalitions between political parties,establishing the identity of the incumbent can be complex in some cases. For example, in 2010the incumbent mayor of the municipality of Tuxtla Gutierrez (in the state of Chiapas) representeda four-party coalition containing the PRD, PT, PVEM and PC. However, the 2010 election sawfive groups stand for election: while the PRI, PT, PVEM and PSD all stood separately, a coalition

26These are the only four responses to an open-ended question that received a non-negligiblenumber of mentions.

27Since turnout is not significantly affected by the release of audit reports, we obtain similarresults when measuring vote share as a proportion of registered voters.

11

formed between the PAN, PRD, PC and PANAL. To address such cases where the incumbentcoalition split at the next election, we define the incumbent vote share as the vote share of thelargest sub-coalition at the next municipal election, and define incumbent re-election similarly.This captures the fact that many coalitions are dominated by a major party in the municipality.Although most cases are straight-forward to code, we ensure that our results do not depend on ourcoding by restricting our analysis to the 82% of electoral precincts in our sample where the PANor PRI—Mexico’s largest national parties, which typically dominate every coalition they are partof—were incumbents as a robustness check.28

3.2 Audit reports

Audit reports are released with a two year lag, so the release coincides with elections held twoyears after the report occurred. Since all reports are released in February, and municipal electionstake place later in the calendar year, we simply define elections by whether an audit was releasedin the February of that year. The two year gap between audit and its release means that the auditgenerally occurred in the first year of a mayor’s term in the cases where the report was released inan election year. Our control group will be mayors in municipalities where the audit was releasedin the year following the election; in such cases, the audit report generally pertains to their secondyear in office.

The results of audit reports are publicly available on the ASF’s website. Each report providesspecific quantitative details pertaining to the use of FISM funds. We extracted the proportion offunds spend in an unauthorized manner and not spend on project benefiting the poor from everyavailable report between 2005 and 2012.29 This yielded a total 1,050 municipal audits, whichwere relatively evenly spread across years and covered 432 unique municipalities. Of these (andthose for which data is available), 388 reports from 268 different municipalities were released in anelection year or the year after. We henceforth restrict attention to this subsample of audits, whichare depicted in Figure 3.

[Figure 3 about here.]

We focus on two measures of mayor performance. To capture corruption, we define two indi-cators for precincts with mayors in the third and fourth quartiles of the distribution of unauthorized

28Although the PRD is also a large national party, its support base is more regional and conse-quently often forms coalitions with parties that are locally strong.

29We did not collect earlier audit reports because they did not present those figures in a system-atic way.

12

FISM spending. To capture neglectful spending, we similarly define indicators for mayors in thethird and fourth quartiles with respect to FISM funds not allocated to spending on the poor. Inour sample, only mayors above the median engaged in non-negligible corrupt or neglectful spend-ing.30 We prefer binary performance metrics identifying more egregious cases of bad performancesince standard theoretical models suggest that voter sanctioning involves cut-off rules (e.g. Barro1973; Ferejohn 1986). Furthermore, our examination of media reports indicates that only relativelyserious cases are widely reported. Nevertheless, we find very similar results using a continuousmeasure where we instead assume that sanctioning is a linear function of revealed performance.

3.3 Media coverage

In addition to our fine-grained electoral data, a key feature of this study is the detail of our mediacoverage data. Following a major media reform in 2007 (see Larreguy, Marshall and Snyder Jr.2014), the IFE required signal coverage data from every AM and FM radio station and everytelevision station in the country. For each media station we code the municipality from where thestation broadcasts and define the commercial quality coverage area.31 See Larreguy, Marshall andSnyder Jr. (2014) for details.

Figures 4, 5, and 6 map the location and coverage of each AM, FM and television station.Although media coverage is extensive, with most precincts receiving at least one media signaland most municipalities containing at least one media station, Figures 7a and 7b below documentconsiderable variation in the number of own-municipality media stations covering each precinct.Since the number of radio and television stations has remained constant between 2003 and 2010,we cannot exploit temporal variation in media coverage.

[Figures 4, 5 and 6 about here.]

Our principal measure of local media coverage is the total number of local media stationscovering a given electoral precinct. We define a local media station as any AM, FM or televisionstation emitting from within a given precinct’s municipality. The average precinct is covered by4.4, 5.4 and 2.4 AM, FM and television stations respectively, while the total number of localmedia stations covering a precinct ranges from 0 to 40. Given these precinct totals are highlycorrelated across media types, simply adding the total together yields similar results to examining

30The level of corruption in the median precinct was 0.4% of FISM funds, while neglectfulspending in the median precinct was 0%.

31For only a small number of FM and television stations did the same station broadcast frommultiple municipalities. No electoral precincts received the same signal from multiple antennae.

13

each type of media separately.32 As a robustness check, we also examine each type of mediaseparately. To identify the importance of local media, we also computed the total number ofmedia stations covering a precinct that do not broadcast from within the municipality. The averageprecinct receives as many FM and television signals from inside their municipality as outside; thegreater signal range of AM stations means that precincts are typically covered by twice as manyAM stations emitting from outside their municipality.

[Figures 7a and 7b about here.]

4 Empirical strategy

Our goal is to identify the effect of local media coverage of municipal audits on incumbent party’selectoral performance. To achieve this, we require exogenous variation in both the release of auditreports and access to local media. We combine the difference-in-difference (DD) design of Ferrazand Finan (2008) with a geographic regression discontinuity (GRD) exploiting within-municipalityvariation in media coverage. We first briefly outline the DD approach used in previous work,before explaining both why the GRD is required to estimate the effect of local media and how it isimplemented.

4.1 Identifying the effects of audit reports

The DD design rests upon exogenous variation in the occurrence of audience. To identify theeffects of audits we use municipalities where an audit report was released just after an electionas a control group for municipalities where the audit was released before the election. We thenmove beyond this first difference by adding two additional layers of differences: whether a mayoris corrupt or neglectful, and the number of media stations covering a given precinct.

As noted above, Mexican municipalities are not randomly chosen for audits. However, wefocus on municipalities that have been audited at least once. Consequently, to identify the causaleffects of an audit being released we require only that the timing of audits is effectively random.Given that the ASF is an independent body and follows a specific formula for allocating audits,this appears to be a reasonable assumption. Table 1 confirms that differences in the political,demographic, media coverage and economic characteristics between electoral precincts in munic-ipalities where an audit was released in the year before an election and those where an audit was

32In our main sample (before exploiting our GRD design), this procedure yielded a Cronbach’salpha of 0.84. The minimum pairwise correlation between the variables is 0.65.

14

released the following year are consistent with chance.33 The variables in the final five rows aretaken from the precinct-level Census data from 2010, and are described in greater detail in theOnline Appendix. Our full sample contains 42,595 precinct-election observations.

[Table 1 about here.]

A second potential concern is that the content of audit reports differs across election and non-election years. For example, auditors could be more lenient or more meticulous in the knowledgethat a report will be released in an election year. Alternatively, mayors anticipating the release ofan audit report in an election year may spend more appropriately. To examine these possibilitieswe compare audit reports released in an election to the universe of all other audits that have everbeen carried out in Figure 8. The distribution of unauthorized spending and spending not on thepoor is almost identical. Combined with our randomization check, this strongly suggests that theaudits results released in election years are typical of “normal” auditing.

[Figure 8 about here.]

To identify the effect of revealing a mayor to be corrupt or neglectful before an election, weestimate the following DD equation using OLS:

Yp,m,t = β1auditm,t +β2audit outcomem,t +β3

(auditm,t×audit outcome Q3m,t

)+β4

(auditm,t×audit outcome Q4m,t

)+Xp,m,tγ + ζt + εp,m,t , (1)

where Yp,m,t is the incumbent party’s vote share in precinct p in municipality m in year t (or whetherthe incumbent party won the municipal election), auditm,t as an indicator for an audit being releasedbefore the election, and audit outcome Q3m,t and audit outcome Q4m,t are indicators for municipal-ities in the third and fourth quartiles of the distributions corrupt or neglectful mayors (regardless ofwhether the audit was released before or after the election). We include election year fixed effectsζt to ensure that we compare municipalities where an audit report was released just after an elec-tion to municipalities where the audit was released before the election. To increase the efficiencyof our estimates, we also included the political and Census variables listed in Table 1 as controls.Throughout, we cluster by municipality-year to account spatial correlation of the precincts in thesame municipality.

33This assumes that the tests are independent. Although they probably are not, finding only three(of 30) differences statistically significant at the 10% level strongly supports our claim.

15

Our main coefficients of interest are β3 and β4, which identify the effect of an audit conditionalupon it revealing corruption or that the mayor did not spend FISM money on the poor. By notweighting our observations, our estimates reflect the effect of an audit in an average precinct. Thiscontrasts with Ferraz and Finan (2008), whose unit is the average municipality.

4.2 Identifying the effects of media stations revealing audit reports

To examine the heterogeneous effects of revealing corruption or neglectful behavior, Ferraz andFinan (2008) further interact auditm,t × audit outcomem,t with the number of local media stationslocated in a municipality. If the number of local media stations were effectively randomly assigned,then this would estimate the average effect of an audit report being released for each additionallocal media station. Unlike our precinct-level data, this strategy rests upon between-municipalitydifferences in media coverage.

However, media stations are not randomly assigned across municipalities. Although the num-ber of local media stations does not predict political or demographic variables in our data, it issignificantly positively correlated with precinct-level literacy rates, living in households with ba-sic necessities and luxury amenities, and the number of household owning a radio or television(even after the inclusion of municipality and year fixed effects). These correlations may upwardlybias our estimates of local media’s effects if the better educated and informed citizens in suchprecincts are more willing or able to sanction incumbent mayors (e.g. Alt, Lassen and Marshall2014; Weitz-Shapiro and Winters 2014).34 This problem may be worsened by the omission ofadditional unobserved correlates of local media stations. To address these concerns we also exploitplausibly exogenous variation in the number of local media stations.

4.3 Geographic regression discontinuity

To generate plausibly exogenous variation in local media coverage, we compare neighboring elec-toral precincts either side of a media coverage boundary. It is important to emphasize that broadcastsignals decay gradually rather than abruptly, so the commercial quality signal boundary is not sim-ply the difference between receiving or not receiving a station’s signal. However, since our unitof analysis—the electoral precinct—is spatially discrete, any continuously declining differenceproduces a discrete change in signal quality once aggregated to the precinct level. In this sense,we regard the boundary as a fuzzy discontinuity where the likelihood that the average voter in a

34In theory, these correlations could also downwardly bias our estimates if such precincts containvoters with stronger prior belief about their incumbent’s quality (Zaller 1992).

16

precinct receives a high quality signal is substantially lower. Our geographic regression disconti-nuity (GRD) design is therefore similar to previous studies exploiting differences in media marketboundaries (e.g. Ansolabehere, Snowberg and Snyder 2006; Enikolopov, Petrova and Zhuravskaya2011; Snyder and Stromberg 2010).

After identifying “treated” precincts which differ from their neighbors in terms of the numberof local media stations that they receive, we use matching to identify the best neighboring controlprecinct for each treated precinct. The number of local media stations is already well balancedacross political and demographic variables, so our best match is defined as the neighboring precinctwith the smallest Malhanobis distance in terms of the five 2010 Census variables at the foot ofTable 1. Since the Census variables are time-invariant, matched pairs are identical for each year.Together, this yields a GRD sample size of 19,230 observations.

[Table 2 about here.]

Our design ensures that within-match differences in the number of local media stations cov-ering neighboring electoral precincts are plausibly exogenous. Table 2 shows that the number oflocal media stations is generally uncorrelated with audit, demographic and economic variables inregressions controlling for match and year fixed effects.35 Importantly, there are no significantdifferences in the number of non-local media stations; this is important because the number oflocal media stations is typically correlated with non-local media stations.36 There are, however,two cases of slight imbalance. Driven by a small number of cross-municipalities matches, PANmunicipalities have more media stations than PRI and then PRD municipalities. Internet usage isalso 0.1 percentage points higher for an additional local media station. In our robustness checks,we demonstrate that neither represents a source of concern.

The GRD sample differs from the full sample of elections occurring in the years before andafter an audit is released. Consequently, our sample may represent different populations, and theeffectively random assignment of audits may no longer hold. Table 3 provides summary statisticsfor the GRD sample, and compares municipalities by the timing of their audit’s release. First,corruption is six percentage points less prevalent in the GRD sample, while media coverage isslightly larger, and precincts are generally less socio-economically developed. Second, there areno significant differences between precincts in municipalities audited before and after electionswhen comparing precincts where an audit was released before and after an election.

35The media variables are omitted because they define the matching design.36In the Online Appendix, we also show that simultaneously controlling for the number of non-

local media stations does not affect balance across the number of local media stations.

17

[Table 3 about here.]

Combining exogenous variation in the timing of audit and the number of media stations cover-ing a given electoral precinct, we estimate the following DD and GRD specification:

Yp,k,m,t = β1auditm,t +β2audit outcomem,t +β3

(auditm,t×audit outcome Q3m,t

)+β4

(auditm,t×audit outcome Q4m,t

)+β5mediap,m +β6

(auditm,t×mediap,m

)+β7

(auditm,t×audit outcome Q3m,t×mediap,m

)+β8

(auditm,t×audit outcome Q4m,t×mediap,m

)+Xp,m,tγ + ξk + ζt + εp,k,m,t , (2)

where ξk is a match fixed effect, which ensures our estimates identify only off within-match vari-ation in media coverage. The balancing variables in Table 5 are included in Xp,m,t to increasethe efficiency of our estimates. To understand the effects of media, we examine two measuresof mediap,m: the total number of local media stations and the total number of non-local mediastations. Since our sample contains some precincts at municipality borders to maximize our sam-ple size, we exploit both within- and across-municipality variation. However, if we only allowwithin-municipality matches or include municipality fixed effects, we obtain very similar results.

5 Results

We first examine whether audits revealing a mayor to be corrupt or neglectful before an electionreduce a mayor’s vote share and probability of re-election. However, our main contribution is tothen identify large complementary effects of local media. Our results demonstrate that an addi-tional media station significantly increases the electoral punishment that a corrupt or neglectfulmayor faces.

5.1 Audits and political accountability

Table 4 presents the average effect of an audit revealing a mayor to be corrupt or neglectful of thepoor on the mayor’s electoral prospects. The outcome in columns (1) and (2) is the change in themayor’s vote share at the precinct level, while the outcome in columns (3) and (4) is whether themayor was re-elected in the municipality. The results pertain to the full sample, and were estimatedusing equation (1).

[Table 4 about here.]

18

Our estimates indicate that an audit released before an election can have substantial electoralimplications. Column (1) shows that revealing a mayor to be in the most corrupt quartile beforethe election, on average, reduces the vote share of a corrupt mayor by seven percentage points.Although this does not quite reach statistical significance, the magnitude represents 14% of theaverage incumbent’s initial vote share. Column (2) finds that revealing that a mayor is neglectfulbefore an election significantly reduces their vote share: the vote share of mayors in the thirdquartile declines by eight percentage points, while mayors in the fourth quartile lose a further threepercentage points.

Looking at the probability of re-election similarly suggests that voters severely punish mayoralmalfeasance. Measured at the municipal level, the change in incumbent vote share maps to largereductions in the probability of being re-elected. Column (3) finds that revealing a mayor as oneof the most corrupt reduces their re-election probability by 34 percentage points, although this isagain not quite statistically significant.37 Column (4) shows that the publication of an audit reportshowing that a mayor did not spend FISM federal transfers on the poor is 38 percentage points lesslikely to be re-elected. For both audit outcomes, the effect is much larger for mayors in the fourthquartile relative to the third.

The substantial electoral sanctioning implied by these results is similar in magnitude to thatfound by Ferraz and Finan (2008) in Brazil, although we measure corruption in terms of stolenfunds rather than the number of corrupt spending violations. However, our results also suggestthat incorrectly spending money earmarked for the poor evokes sanctioning of similar magnitudeto corruption.38 However, our DD estimates are relatively noisy. A plausible explanation for ourlack of precision is that audit reports only affect voter behavior when the information is clearlyconveyed.

5.2 The effects of media coverage of audit reports

We now address the central question of this article: are politicians more likely to be sanctioned byvoters when media stations exist to publicize their behavior? Combining our DD and GRD designs,we first examine the effects of local media stations—those emitting from the same municipality asan electoral precinct—before turning to non-local media stations. Since mayoral corruption and

37The relative lack of precision reflects the fact that we have 481 audited municipalities, ofwhich only 51 had mayors that were revealed to be corrupt before the election.

38The Online Appendix reports quantitatively similar, but far noisier, estimates for the GRDsample, which includes only a selection of precincts from 330 municipality elections and onlyidentifies off within-matched pair variation.

19

neglect are primarily important local issues, we expect to find that local media is more effective atfacilitating electoral accountability.

Table 5 provides our estimates for the sanctioning effect of an additional media station emittingfrom the precinct’s own municipality. Since the precincts in most matched pairs are both locatedin the same municipality, our analysis now focuses on the change in the incumbent’s vote share atthe precinct level.

The results for revealing a corrupt mayor show that local media supports electoral accountabil-ity. In column (1), mayors in the third and fourth quartiles of the corruption distribution experiencesignificant loses in their vote share—around one percentage point for each additional local mediastation. A standard deviation increase in the number of media stations, which entails 11.6 moremedia stations, thus reduces the vote share of an incumbent revealed to be corrupt by ten per-centage points. The insignificant interactions between our audit dummy and a mayor’s corruptionquartile indicate that revealing a mayor to be corrupt has no effect in precincts covered by no mediastations.

[Table 5 about here.]

In precincts where local media reveals a mayor to have neglected the poor, the effect is largerand increasing in the severity of a mayor’s neglect. Column (2) shows that an additional localmedia station reduces a neglectful mayor’s vote share by 1.5 percentage points for mayors in thethird quartile and 2.5 percentage points for mayors in the most neglectful quartile. A standarddeviation increase in the number of local media stations thus entails a 23 percentage point decreasein the vote share of the most neglectful mayors if their behavior is revealed before an election.This represents a decline of almost half their vote share. Again, the interaction between the pre-election audit release and not spending on the poor shows that in locations with zero local mediastations, the effect of revealing a mayor to be neglectful is effectively zero. Consistent with theDD estimates in Table 4), our GRD results thus clearly indicate that the media-induced electoralresponse to revealing neglect of the poor is larger than the electoral response to corruption.

To examine the role of non-local media, Table 6 presents the results of estimating equation(2) with interactions for both local and non-local media. For non-local media, balance acrossthe variables in Table 2 is even stronger.39 The results clearly indicate that non-local media hasa far smaller effect on election outcomes than local media. Comparing the coefficient on thetriple interactions, the effect of non-local media is always smaller than for local media. While the

39Since our matching did not maximize differences in non-local media, this in part stems frommany matches with no variation in the number of non-local media stations.

20

difference for corruption is relatively small, the large impact of local media is far greater than thatfor non-local media where a mayor neglects the poor. Nevertheless, the negative coefficients fornon-local media allow for the possibility of a small spillover effect consistent with media stationsreporting audit results reports for multiple local municipalities.40

[Table 6 about here.]

5.3 Robustness checks

By exploiting two sources of plausibly exogenous variation, there are good reasons to be confidentin our estimates. Nevertheless we now show that our GRD estimates are not sensitive to particularspecification choices. Table 8 present the results of our robustness checks, and focus on the tripleinteractions identifying the effect of local media revealing mayoral malfeasance.

First, we show that the results are robust to our definition of party incumbents. Since partycoalitions can vary over time within municipalities, and it is not always obvious which the promi-nent party within a given coalition is, we restrict attention to incumbents containing the PAN orPRI. These parties almost always dominate smaller parties in local coalitions. Panel A of Table 8shows that our estimates slightly increase in magnitude for the 88% of precincts where the PAN orPRI were incumbents.

[Table 8 about here.]

Second, panels B-D show the results when looking at different types of media separately. Com-paring coefficients magnitudes across the panels, we find that local media’s effects are generallylargest for FM radio and especially television. There are several potential explanations for thisfinding. First, as the coverage maps above indicate, AM radio stations cover a larger area and maythus seek to appeal to a broader demographic. Second, the matches that differ in terms of their FMand television coverage may be more urban locations where voters are more able or willing to actupon information concerning their mayor’s performance.

Third, a potential concern is that our GRD estimates are driven by matches where the treatedand control units are from bordering municipalities. Such matches could therefore be pickingup unobservable features of the municipalities, or the lingering differences in incumbent statushighlighted in Table 2. Demonstrating that this is not the case, panel E provides very similar

40Because our GRD-matching approach focuses on difference in local media, our estimates fornon-local media may be under-powered.

21

estimates when controlling for municipality fixed effects. Panel F shows very similar results whenthe set of potential matches is restricted to precincts within the same municipality.

Fourth, our balance checks highlighted an imbalance in internet access. Given audit reportswere also released online, it is possible that local media coverage is simply a proxy for internetaccess. However, since the imbalance is extremely small in magnitude, it is hard to believe that thiscorrelation could account for our large estimates. Furthermore, panel G shows that controlling for atriple interaction between audit results and the proportion of the precinct with internet access barelyaffects our point estimates, although the corruption coefficients become statistically insignificant.Supporting the importance of media more generally, the internet coefficients suggest that precinctswith greater internet access are more likely to punish corrupt, but not neglectful, mayors. Sinceinternet access may still be correlated with precinct characteristics like education and politicalinterest, further work is required to establish whether the association with internet access is causal.

Fifth, we consider a linear specification of the audit report results. In particular, we use theshare of unauthorized spending and spending not on the poor instead of the binary approachesused above. We thus allow for alternative types of voter punishment strategies. Panel H showssubstantively similar results, although the magnitudes ares smaller. This supports the claim thatvoters more severely punish worse behavior in office.

Sixth, at the cost of losing randomization in local media, we estimate equation (1) on the fullsample in order to check the external validity of our GRD estimates. The results in Table 9 arebroadly similar to the GRD estimates. For both corruption and spending not on the poor, a mayorrevealed as malfeasant before an election experiences a significant decrease in their vote share foreach additional local media station. Although the GRD and full samples yield relatively similarresults, the estimates suggest that not accounting for differences correlated with media availabilityslightly downwardly biases media’s amplifying effects of spending not on the poor.

[Table 9 about here.]

5.4 Heterogeneity by party

Since parties vary in their political platforms and reputations, voters might punish some partiesmore than others. For example, voters might be more likely to update their beliefs about the sin-cerity, commitment, or competence of parties when the incumbent’s behavior in office contradictsthe campaign promises of their party or voter expectations of politicians.41 Alternatively, voters

41See, e.g. Alt, Lassen and Marshall (2014), who show that opposition claims that economyis performing well or government claims that the economy is performing badly have larger ef-

22

might find some behavior particularly hypocritical, and therefore more egregious and deserving ofpunishment.

While all parties in Mexico have been linked with corruption, particularly in our samplingperiod following the collapse of PRI single-party dominance, the main parties differ in the extentto which they promise to support the poor. Since the PRI is a populist party appealing to therelatively disadvantaged masses, we might expect revelations of failing to spend no the poor tohurt the PRI more than other parties. The PAN, which also has a significant number of mayors, isinstead the party of richer voters.

Table 10 examines this relationship by comparing the effects of media in the GRD sample be-tween PRI and non-PRI incumbents.42 Although PRI municipalities could be correlated with otherrelevant characteristics, the results suggest that our results are primarily driven by the precinctswith a PRI incumbent. This evidence implies that voters are particularly willing to punish politi-cians that claim to support poorer voters but which ultimately neglect them.

[Table 10 about here.]

6 Conclusion

Many scholars call media “the fourth estate,” due to its potential to inform voters about the behaviorof politicians in office. Both national and local media are needed: while national media outletscover national level actors, local media are necessary to inform voters about the performance oflocal politicians. However, since local media is often monopolistic or oligopolistic, it may beespecially vulnerable to capture in developing democracies (Besley and Prat 2006).

Using detailed local data and a geographical regression discontinuity design that exploits dif-ferences in signal coverage across neighboring electoral precincts, we identify the impact of themedia environment on political accountability. We show that voters punish the party of malfeasantmayors, but only in electoral precincts covered by local media stations. In particular, we find thateach additional local radio or television station reduces the vote share of an incumbent politicalparty revealed to be corrupt by 1 percentage point, and reduces the vote share of an incumbent po-litical party revealed to be have diverted funds away from the poor by about 2.5 percentage points.

fects on voter evaluations of government. Similarly, Chiang and Knight (2011) for evidence that“surprising” newspaper endorsements have a significant effect on voter behavior but “expected”endorsements do not.

42Although the PRD is perhaps the most left-leaning party, the number of PRD mayors is quitesmall and these mayors often hold office in coalition with other local parties.

23

However, we find no effect of media stations that cover the municipality but are based in othermunicipalities. Thus, our findings demonstrate the importance of media, especially local media, insupporting political accountability.

The electoral costs of diverting resources away from the poor are especially large for the pop-ulist PRI party. One interpretation of this finding is that voters punish parties more for behaviorthat is not only malfeasant but contrary to the party’s ideological reputation. This raises importantquestions about voter sophistication and, therefore, about the scope for politicians to engage inmalfeasant behavior in office. Further work is clearly needed to understand the conditions underwhich different types of politicians are punished by voters.

24

References

Alt, James E., David D. Lassen and John Marshall. 2014. “Information sources, belief updating,and the politics of economic expectations: Evidence from a Danish survey experiment.” Workingpaper.

Ansolabehere, Stephen, Erik C. Snowberg and James M. Snyder. 2006. “Television and the incum-bency advantage in US elections.” Legislative Studies Quarterly 31(4):469–490.

Banerjee, Abhijit V., Donald Green, Jeffrey McManus and Rohini Pande. 2014. “Are Poor VotersIndifferent to Whether Elected Leaders are Criminal or Corrupt? A Vignette Experiment inRural India.”. Working paper.

Banerjee, Abhijit V., Selvan Kumar, Rohini Pande and Felix Su. 2011. “Do Informed Voters MakeBetter Choices? Experimental Evidence from Urban India.” Working paper.

Barro, Robert J. 1973. “The control of politicians: an economic model.” Public choice 14(1):19–42.

Besley, Timothy and Andrea Prat. 2006. “Handcuffs for the grabbing hand? Media capture andgovernment accountability.” American Economic Review pp. 720–736.

Boas, Taylor C. and F. Daniel Hidalgo. 2011. “Controlling the airwaves: Incumbency advantageand community radio in Brazil.” American Journal of Political Science 55(4):869–885.

Bobonis, Gustavo J., Luis R. Camara Fuertes and Rainer Schwabe. 2013. “Monitoring CorruptiblePoliticians.” Working paper.

Bobonis, Gustavo J., Luis R. Camara Fuertes and Rainer Schwabe. 2014. “Monitoring CorruptiblePoliticians.” Working paper.

Caselli, Francesco and Massimo Morelli. 2004. “Bad politicians.” Journal of Public Economics

88(3):759–782.

Chang, Eric C.C., Miriam A. Golden and Seth J. Hill. 2010. “Legislative malfeasance and politicalaccountability.” World Politics 62(2):177–220.

Chiang, Chun-Fang and Brian Knight. 2011. “Media bias and influence: Evidence from newspaperendorsements.” Review of Economic Studies 78(3):795–820.

25

Chong, Alberto, Ana De La O, Dean Karlan and Leonard Wantchekon. forthcoming. “Does Cor-ruption Information Inspire the Fight or Quash the Hope? A Field Experiment in Mexico onVoter Turnout, Choice and Party Identification.” Journal of Politics .

Costas, Elena, Albert Sole-Olle and Pilar Sorribas-Navarro. 2011. “Do voters really tolerate cor-ruption? Evidence from Spanish Mayors.” Working paper.

de Figueiredo, Miguel F.P., F. Daniel Hidalgo and Yuri Kasahara. 2013. “When Do Voters PunishCorrupt Politicians? Experimental Evidence from Brazil.” Working paper.

DellaVigna, Stefano and Ethan Kaplan. 2007. “The Fox News effect: Media bias and voting.”Quarterly Journal of Economics 122(3):1187–1234.

DellaVigna, Stefano and Matthew Gentzkow. 2010. “Persuasion: Empirical Evidence.” Annual

Review of Economics 2(1):643–669.

Dutta, Bhaskar and Poonam Gupta. 2012. “How Do Indian Voters Respond to Candidates withCriminal Charges: Evidence from the 2009 Lok Sabha Elections.” University Library of Munich,MPRA Paper Series, No. 38417.

Eggers, Andrew and Alexander C. Fisher. 2011. “Electoral Accountability and the UK Parliamen-tary Expenses Scandal: Did Voters Punish Corrupt MPs?” SSRN working paper.

Enikolopov, Ruben, Maria Petrova and Ekaterina Zhuravskaya. 2011. “Media and political per-suasion: Evidence from Russia.” American Economic Review 101(7):3253–3285.

Ferejohn, John. 1986. “Incumbent performance and electoral control.” Public choice 50(1):5–25.

Fergusson, Leopoldo. 2014. “Media markets, special interests, and voters.” Journal of Public

Economics 109(C):13–26.

Ferraz, Claudio and Frederico Finan. 2008. “Exposing Corrupt Politicians: The Effects of Brazil’sPublicly Released Audits on Electoral Outcomes.” Quarterly Journal of Economics 123(2):703–745.

Humphreys, Macartan and Jeremy Weinstein. 2012. “Policing Politicians: Citizen Empowermentand Political Accountability in Uganda Preliminary Analysis.” Working paper.

Klasnja, Marko. 2011. “Why Do Malfeasant Politicians Maintain Political Support? Testing the“Uninformed Voter” Argument.”.

26

Larreguy, Horacio, John Marshall and James M. Snyder Jr. 2014. “PPolitical Advertising in Con-solidating Democracies: Dominant Parties and the Equalization of Media Access in Mexico.”Working paper.

Lavielle, Briseida, Kristina Pirker and Alberto Serdan. 2006. “Monitoreo de Programas Socialesen Contextos Electorales.” Fundar Working paper.

McCann, James A. and Jorge I. Domınguez. 1998. “Mexicans react to electoral fraud and politicalcorruption: an assessment of public opinion and voting behavior.” Electoral Studies 17(4):483–503.

Niehaus, Paul and Sandip Sukhtankar. 2013. “Corruption Dynamics: The Golden Goose Effect.”American Economic Journal: Economic Policy 5(4):230–269.

Olken, Benjamin A. 2007. “Monitoring Corruption: Evidence from a Field Experiment in Indone-sia.” Journal of Political Economy 115(2).

Rundquist, Barry S., Gerald S. Strom and John G. Peters. 1977. “Corrupt politicians and theirelectoral support: some experimental observations.” American Political Science Review pp. 954–963.

Slomczynski, Kazimierz and Goldie Shabad. 2012. “Perceptions of political party corruption andvoting behaviour in Poland.” Party Politics 18(6):897–917.

Snyder, James M. and David Stromberg. 2010. “Press Coverage and Political Accountability.”Journal of Political Economy 118(2):355–408.

Stanig, Piero. forthcoming. “Regulation of Speech and Media Coverage of Corruption: An Empir-ical Analysis of the Mexican Press.” American Journal of Political Science .

Svolik, Milan W. 2013. “Learning to love democracy: electoral accountability and the success ofdemocracy.” American Journal of Political Science 57(3):685–702.

Weitz-Shapiro, Rebecca and Matthew S. Winters. 2014. “Discerning corruption: Credible Accu-sations and the Punishment of Politicians in Brazil.” Working paper.

Wellenstein, Anna, Angelica Nunez and Luis Andres. 2006. Social Infrastructure: Fondo deAportaciones para la Infraestructura Social (FAIS). In Decentralized service delivery for the

poor, Volume II: Background papers, ed. The World Bank. Mexico City: The World Bankpp. 167–222.

27

-2-1

01

23

Sta

ndar

dize

d se

arch

act

ivity

(si

gma)

Feb 06 Feb 07 Feb 08 Feb 09 Feb 10 Feb 11 Feb 12

Figure 1: Google searches for ASF, by month

Notes: Extracted using Google Correlate (http://correlate.googlelabs.com) on 15th July 2014. The datacover the period used in our sample.

Zaller, John R. 1992. The Nature and Origins of Mass Opinion. Cambridge University Press.

28

-10

12

3S

tand

ardi

zed

sear

ch a

ctiv

ity (

sigm

a)

Feb 06Feb 06 Feb 07Feb 06 Feb 07 Feb 08Feb 06 Feb 07 Feb 08 Feb 09Feb 06 Feb 07 Feb 08 Feb 09 Feb 10Feb 06 Feb 07 Feb 08 Feb 09 Feb 10 Feb 11Feb 06 Feb 07 Feb 08 Feb 09 Feb 10 Feb 11 Feb 12

Figure 2: Google searches for FISM, by month

Notes: Extracted using Google Correlate (http://correlate.googlelabs.com) on 15th July 2014. The datacover the period used in our sample.

29

Figure 3: Distribution of audit report outcomes by municipality.

Notes: Only the 268 municipalities in our final sample are included. Where more than one audit occurs,we take the average audit outcome.

30

Figure 4: AM radio signal coverage areas (source: IFE).

Figure 5: FM radio signal coverage (source: IFE).

31

Figure 6: TV signal coverage (source: IFE).

32

(a) Whole country.

(b) Zoomed in at the center of the country.

Figure 7: Distribution of the number of own municipality stations (AM, FM and TV) received byprecincts (source: IFE).

Notes: Only precincts in our GRD sample are included. Dark shades of red indicate precincts with alarger total number of own-municipality radio or television stations.

33

05

1015

Den

sity

0 .2 .4 .6 .8 1

Proportion of spending unauthorized

05

1015

Den

sity

0 .2 .4 .6 .8 1

Proportion of spending not on the poor

Non-election Election

Figure 8: Distribution of audit report results

Notes: The distributions are based on 236 audits released in election years, and 506 that were not.

34

Table 1: Summary statistics by audit status (full sample)

Control (no audit) mean Audit difference

Unauthorized spending 0.113 -0.037 (0.031)Corrupt Q3 0.213 0.063 (0.072)Corrupt Q4 0.319 -0.117 (0.081)Spending not on the poor 0.094 0.028 (0.034)Not poor Q3 0.094 -0.031 (0.080)Not poor Q4 0.270 0.026 (0.080)Registered voters 1,340.4 -54.3 (82.82)Turnout (lag) 0.483 0.035* (0.020)PAN incumbent 0.374 0.026 (0.089)PRI incumbent 0.451 0.062 (0.091)PRD incumbent 0.159 -0.072 (0.056)Local AM 0.719 0.019 (0.050)Local FM 0.628 0.047 (0.078)Local TV 0.495 0.157* (0.08)Non-local AM 0.158 0.004 (0.047)Non-local FM 0.235 -0.029 (0.072)Non-local TV 0.349 -0.137* (0.075)Total local AM 3.854 0.607 (0.971)Total local FM 4.405 1.926 (1.239)Total local TV 2.060 0.618 (0.064)Total local media 10.318 3.151 (2.465)Total non-local AM 21.174 -2.006 (3.597)Total non-local FM 10.475 -0.331 (1.903)Total non-local TV 7.315 -1.036 (1.846)Total non-local media 3.383 -0.639 (0.74)Share employed 0.952 0.002 (0.002)Share illiterate 0.060 -0.002 (0.010)Share with household necessities 0.083 0.028 (0.024)Share with household amenities 0.548 0.021 (0.017)Share with internet 0.224 0.019 (0.019)

Notes: The audit difference results are from a regressions of the outcome variables on the left-hand-sideof the table on an indicator for an audit being released the year before an election, where standard errorsare clustered by municipality-year are in parentheses. There are 42,595 observations for each variable.* denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

35

Tabl

e2:

Bal

ance

afte

rGR

D-m

atch

ing

Pane

lA:T

ime-

vary

ing

vari

able

sU

naut

h.C

orru

ptN

on-p

oor

Not

poor

Elig

ible

Turn

out

PAN

PRI

PRD

spen

ding

spen

ding

vote

rs(l

ag)

incu

mbe

ntin

cum

bent

incu

mbe

nt

Loc

alm

edia

-0.0

01-0

.000

-0.0

01-0

.001

15.8

07-0

.001

0.02

2***

-0.0

08-0

.015

***

(0.0

03)

(0.0

08)

(0.0

04)

(0.0

04)

(14.

115)

(0.0

01)

(0.0

07)

(0.0

06)

(0.0

06)

Obs

erva

tions

19,2

3019

,230

19,2

3019

,230

19,2

3019

,230

19,2

3019

,230

19,2

30

Pane

lB:T

ime-

inva

rian

tvar

iabl

esN

on-l

ocal

Em

ploy

edIl

liter

ate

Hou

seho

ldH

ouse

hold

Inte

rnet

med

iane

cess

ities

amen

ities

Loc

alm

edia

-0.0

410.

000

-0.0

00-0

.000

0.00

10.

001*

*(0

.146

)(0

.000

)(0

.000

)(0

.001

)(0

.001

)(0

.001

)

Obs

erva

tions

19,2

3019

,230

19,2

3019

,230

19,2

3019

,230

Not

es:

All

spec

ifica

tions

incl

ude

mat

chan

dye

arfix

edef

fect

s,up

totw

opo

ssib

lem

atch

es,a

ndar

ees

timat

edus

ing

OL

S.St

anda

rder

rors

are

clus

tere

dby

mun

icip

ality

-yea

r.*

deno

tes

p<

0.1,

**de

note

sp<

0.05

,***

deno

tes

p<

0.01

.

36

Table 3: Summary statistics by audit status (GRD sample)

Control (no audit) mean Audit difference

Unauthorized spending 0.062 0.006 (0.023)Corrupt Q3 0.243 0.099 (0.095)Corrupt Q4 0.186 -0.001 (0.067)Spending not on the poor 0.064 0.043 (0.029)Not poor Q3 0.335 -0.036 (0.108)Not poor Q4 0.113 0.132** (0.066)Registered voters 1,380.9 26.21 (153.9)Turnout (lag) 0.493 0.016 (0.020)PAN incumbent 0.395 -0.068 (0.101)PRI incumbent 0.457 0.115 (0.102)PRD incumbent 0.117 -0.01 (0.043)Local AM 0.817 0.011 (0.050)Local FM 0.718 0.024 (0.059)Local TV 0.580 0.072 (0.065)Non-local AM 0.135 -0.026 (0.038)Non-local FM 0.139 -0.029 (0.038)Non-local TV 0.192 -0.071* (0.038)Total local AM 4.071 0.720 (1.085)Total local FM 4.242 2.185* (1.304)Total local TV 2.313 0.308 (0.064)Total local media 10.626 3.213 (2.845)Total non-local AM 12.808 -0.572 (2.223)Total non-local FM 8.122 -0.68 (1.314)Total non-local TV 3.118 0.303 (0.853)Total non-local media 1.569 -0.195 (0.355)Share employed 0.954 0.002 (0.003)Share illiterate 0.092 -0.008 (0.010)Share with household necessities 0.666 0.059 (0.036)Share with household amenities 0.466 0.041 (0.026)Share with internet 0.127 0.043 (0.027)

Notes: The audit difference results are from a regressions of the outcome variables on the left-hand-sideof the table on an indicator for an audit being released the year before an election, where standard errorsare clustered by municipality-year are in parentheses. There are 19,230 observations for each variable.* denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

37

Table 4: The effect of audits revealing corruption (full sample)

Change in incumbent vote share Incumbent re-elected(1) (2) (3) (4)

Audit 0.028 0.062*** 0.122 0.147(0.021) (0.024) (0.120) (0.112)

Corrupt Q3 0.009 0.070(0.027) (0.143)

Audit × Corrupt Q3 -0.001 0.046(0.041) (0.175)

Corrupt Q4 0.028 0.278*(0.041) (0.157)

Audit × Corrupt Q4 -0.066 -0.337(0.046) (0.215)

Not poor Q3 0.043* -0.072(0.026) (0.156)

Audit × Not poor Q3 -0.080* 0.013(0.041) (0.189)

Not poor Q4 0.076* 0.100(0.040) (0.152)

Audit × Not poor Q4 -0.111** -0.377**(0.046) (0.187)

Observations 42,595 42,595 42,595 42,595R2 0.08 0.09 0.14 0.14

Notes: All specifications include pre-treatment controls and year fixed effects, and are estimated usingOLS. Similar estimates for the GRD sample are provided in the Online Appendix. The omitted categoryfor corruption and not spending on the poor is Q1 and Q2. Standard errors are clustered by municipality-year. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

38

Table 5: The effect of audits revealing corruption, by local media availability (GRD sample)

Change in incumbent vote share(1) (2)

Audit -0.039 -0.013(0.034) (0.032)

Local media -0.005 -0.007*(0.004) (0.004)

Audit × Local media 0.006* 0.012***(0.003) (0.004)

Corrupt Q3 -0.105**(0.046)

Audit × Corrupt Q3 0.078(0.053)

Audit × Corrupt Q3 × Local media -0.008**(0.004)

Corrupt Q4 0.001(0.057)

Audit × Corrupt Q4 0.064(0.064)

Audit × Corrupt Q4 × Local media -0.009(0.006)

Not poor Q3 -0.018(0.039)

Audit × Not poor Q3 0.043(0.047)

Audit × Not poor Q3 × Local media -0.015***(0.005)

Not poor Q4 -0.036(0.045)

Audit × Not poor Q4 0.035(0.072)

Audit × Not poor Q4 × Local media -0.025***(0.007)

Observations 19,230 19,230R2 0.52 0.54

Notes: All specifications include pre-treatment controls and match and year fixed effects as controls,use up to two possible matches, and are estimated using OLS. The omitted category for corruption andnot spending on the poor is Q1 and Q2. Standard errors are clustered by municipality-year. * denotesp < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

39

Table 6: The effect of audits revealing corruption, including non-local media (GRD sample)

Change in incumbent vote share(1) (2)

Audit -0.067 0.020(0.045) (0.035)

Audit × Corrupt Q3 0.146**(0.063)

Audit × Corrupt Q3 × Local media -0.007**(0.004)

Audit × Corrupt Q3 × Non-local media -0.005**(0.003)

Audit × Corrupt Q4 0.117(0.076)

Audit × Corrupt Q4 × Local media -0.009(0.006)

Audit × Corrupt Q4 × Non-local media -0.003(0.003)

Audit × Not poor Q3 × Local media -0.015***(0.005)

Audit × Not poor Q3 × Non-local media 0.005*(0.003)

Audit × Not poor Q4 0.069(0.089)

Audit × Not poor Q4 × Local media -0.020***(0.006)

Audit × Not poor Q4 × Non-local media -0.001(0.003)

Observations 19,230 19,230R2 0.53 0.54Equality of media effects (p value)

Corrupt Q3 0.649Corrupt Q4 0.429Not poor Q3 0.000Not poor Q4 0.007

Notes: All specifications include pre-treatment controls and match and year fixed effects, use up to twopossible matches, and are estimated using OLS. The omitted category for corruption and not spending onthe poor is Q1 and Q2. Some lower order terms are omitted to save space. Standard errors are clusteredby municipality-year. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

40

Table 7: Robustness checks (GRD sample)

Panel A: PAN and PRI incumbents Change in incumbent vote share(1) (2)

Audit × Corrupt Q3 × Local media -0.009**(0.004)

Audit × Corrupt Q4 × Local media -0.013**(0.006)

Audit × Not poor Q3 × Local media -0.013**(0.006)

Audit × Not poor Q4 × Local media -0.032***(0.007)

Panel B: AM radio Change in incumbent vote share(1) (2)

Audit × Local AM × Corrupt Q3 0.008(0.011)

Audit × Local AM × Corrupt Q4 0.012(0.019)

Audit × Local AM × Not poor Q3 -0.029(0.019)

Audit × Local AM × Not poor Q4 -0.044**(0.019)

Panel C: FM radio Change in incumbent vote share(1) (2)

Audit × Local FM × Corrupt Q3 -0.030***(0.010)

Audit × Local FM × Corrupt Q4 -0.015(0.011)

Audit × Local FM × Not poor Q3 -0.031***(0.011)

Audit × Local FM × Not poor Q4 -0.041***(0.011)

Panel D: television Change in incumbent vote share(1) (2)

Audit × Local TV × Corrupt Q3 -0.021(0.015)

Audit × Local TV × Corrupt Q4 -0.042**(0.016)

Audit × Local TV × Not poor Q3 -0.037***(0.014)

Audit × Local TV × Not poor Q4 -0.062***(0.020)

Notes: All specifications include pre-treatment controls and match and year fixed effects, use up to twopossible matches, and are estimated using OLS. All regressions have 19,230 observations except PanelA which contains 16,947 observations. Standard errors are clustered by municipality-year. * denotesp < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

41

Table 8: Robustness checks (GRD sample) (continued)

Panel E: municipality fixed effects Change in incumbent vote share(1) (2)

Audit × Corrupt Q3 × Local media -0.010**(0.004)

Audit × Corrupt Q4 × Local media -0.011(0.007)

Audit × Not poor Q3 × Local media -0.019***(0.005)

Audit × Not poor Q4 × Local media -0.030***(0.010)

Panel F: within-municipality matches Change in incumbent vote share(1) (2)

Audit × Corrupt Q3 × Local media -0.015***(0.005)

Audit × Corrupt Q4 × Local media -0.013*(0.008)

Audit × Not poor Q3 × Local media -0.020***(0.006)

Audit × Not poor Q4 × Local media -0.027**(0.010)

Panel G: control for internet access Change in incumbent vote share(1) (2)

Audit × Corrupt Q3 × Local media -0.007(0.005)

Audit × Corrupt Q4 × Local media -0.006(0.007)

Audit × Corrupt Q3 × Internet -0.104(0.157)

Audit × Corrupt Q4 × Internet -0.479***(0.158)

Audit × Not poor Q3 × Local media -0.017***(0.005)

Audit × Not poor Q4 × Local media -0.027***(0.007)

Audit × Not poor Q3 × Internet 0.258(0.198)

Audit × Not poor Q4 × Internet 0.097(0.196)

Panel H: linear audit measures Change in incumbent vote share(1) (2)

Audit × Unauthorized × Local media -0.009(0.019)

Audit × Spending not poor × Local media -0.048***(0.014)

Notes: All specifications include pre-treatment controls and match and year fixed effects, use up to twopossible matches, and are estimated using OLS. All regressions have 19,230 observations, except PanelF which contains 15,080 observations. Standard errors are clustered by municipality-year. * denotesp < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

42

Table 9: The effect of audits revealing corruption, by local media availability (full sample)

Change in incumbent vote share(1) (2)

Audit -0.011 -0.001(0.023) (0.025)

Local media 0.000 -0.002(0.001) (0.002)

Audit × Local media 0.001 0.006**(0.001) (0.003)

Corrupt Q3 -0.021(0.031)

Audit × Corrupt Q3 0.017(0.039)

Audit × Corrupt Q3 × Local media -0.002(0.003)

Corrupt Q4 -0.037(0.039)

Audit × Corrupt Q4 0.075(0.055)

Audit × Corrupt Q4 × Local media -0.008**(0.003)

Not poor Q3 0.021(0.031)

Audit × Not poor Q3 -0.018(0.040)

Audit × Not poor Q3 × Local media -0.007**(0.003)

Not poor Q4 -0.010(0.042)

Audit × Not poor Q4 0.009(0.049)

Audit × Not poor Q4 × Local media -0.011***(0.003)

Observations 42,595 42,595R2 0.12 0.14

Notes: All specifications include pre-treatment controls and year fixed effects, and are estimated using OLS.Standard errors are clustered by municipality-year. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

43

Table 10: Heterogeneous effects of revealing neglect of the poor by party

Change in incumbent vote share(1) (2)

Audit -0.013 -0.033(0.032) (0.050)

Local media -0.007* -0.000(0.004) (0.005)

Audit × Local media 0.012*** 0.006(0.004) (0.004)

PRI 0.017(0.056)

Audit × PRI 0.054(0.064)

Local media × PRI -0.003(0.004)

Audit × Local media × PRI 0.007(0.005)

Not poor Q3 -0.018 -0.017(0.039) (0.044)

Audit × Not poor Q3 0.043 0.037(0.047) (0.076)

Audit × Not poor Q3 × Local media -0.015*** -0.004(0.005) (0.007)

Audit × Not poor Q3 × PRI 0.047(0.099)

Audit × Not poor Q3 × Local media × PRI -0.034**(0.014)

Not poor Q4 -0.036 -0.049(0.045) (0.067)

Audit × Not poor Q4 0.035 0.013(0.072) (0.097)

Audit × Not poor Q4 × Local media -0.025*** -0.009(0.007) (0.008)

Audit × Not poor Q4 × PRI 0.058(0.108)

Audit × Not poor Q4 × Local media × PRI -0.027**(0.012)

Observations 19,230 19,230R2 0.54 0.55

Notes: All specifications in the full sample include pre-treatment controls and year fixed effects and are estimatedusing OLS. All specifications in the GRD sample include pre-treatment controls and match and year fixed effects,up to two possible matches, and are estimated using OLS. Lower order media terms are omitted. Standard errorsare clustered by municipality-year. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

44

ONLINE APPENDIX

REVEALING MALFEASANCE: HOW LOCAL MEDIA

FACILITATES ELECTORAL SANCTIONING

OF MAYORS IN MEXICO

HORACIO A. LARREGUY ∗ JOHN MARSHALL †

JAMES M. SNYDER JR. ‡

AUGUST 2014

∗Department of Government, Harvard University, Cambridge, MA (hlarreguy@fas.harvard.edu).†Department of Government, Harvard University, Cambridge, MA (jlmarsh@fas.harvard.edu).‡Department of Government, Harvard University, Cambridge, MA (jsnyder@fas.harvard.edu).

1

1 Variable definitions

Incumbent party vote share. The vote share, as a proportion of total votes cast, in a given electoralprecinct. As noted in the main text, we define incumbent vote share as the vote share of the largestsub-coalition at the next municipal election. We do this because municipal party coalitions changeacross elections. Source: IFE and State Electoral Institutes.Incumbent party win. Indicator coded one where the incumbent party is re-elected. Incumbentsare defined as above. Source: IFE and State Electoral Institutes.Unauthorized spending. Percentage of FISM funds spent in an unauthorized manner. See text forfurther discussion. The variables Corrupt Q3 and Corrupt Q4 are separately defined by the thirdand fourth quartiles of our full sample and GRD sample. Source: ASF audit reports.Spending not on the poor. Percentage of FISM funds spent not spent on the poor. See text forfurther discussion. The variables Not poor Q3 and Not poor Q4 are separately defined by the thirdand fourth quartiles of our full sample and GRD sample. Source: ASF audit reports.Registered voters. The number of voters registered to vote in the electoral precinct. Source: IFEand State Electoral Institutes.Turnout (lag). Precinct-level electoral turnout at the previous election. Source: IFE and StateElectoral Institutes.PAN/PRI/PRD incumbent. Indicator coded one for the incumbent mayor represents a coalitioncontaining the PAN, PRI or PRD. Source: IFE and State Electoral Institutes.Local AM/FM/TV. Indicator coded one for electoral precincts covered by at least one AM/FM/TVstation emitting from within the municipality. Source: computed from IFE data.Non-local AM/FM/TV. Indicator coded one for electoral precincts covered by at least one AM/FM/TVstation emitting from outside the municipality. Source: computed from IFE data.Total local AM/FM/TV. The total number of AM/FM/TV stations emitting from within the munic-ipality. Source: computed from IFE data.Total non-local media. The total number of AM, FM and TV stations emitting from outside themunicipality. Source: computed from IFE data.Share employed. Percentage of the precinct population employed in 2010. Source: Mexican 2010Census.Share illiterate. Percentage of the precinct population aged above 15 that is illiterate in 2010.Source: Mexican 2010 Census.Share with household necessities. Percentage of households with electricity, piped water, toilet anddrainage. Source: Mexican 2010 Census.

2

Share with household amenities. Average number of households in a given precinct with a refriger-ator, a washing machine, a car or truck, a landline, a cellphone, or internet access in 2010. Source:Mexican 2010 Census.Share with internet. Percentage of households in the precinct with an internet connection in 2010.Source: Mexican 2010 Census.

2 Audit reports

Figures 1 and 2 provide an example of an audit report from 2008 for the municipality of Ajalpanin the state of Puebla.

3 Lack of balance across media stations

Table 1 shows that the number of local media stations that cover an electoral precinct is stronglycorrelated with a variety of socio-economic development indicators. In particular, we find thatprecincts with more local media stations are inhabited by household that are more literate, possessmore household necesities and amenities (see variable definitions above), and are more likely toalso be connected to the internet. In short, precincts with more media stations are more developed.

Table 1: Precinct-level correlation of media stations with other socioeconomic variables

Share Share Share with Share with Share withemployed illiterate household household internet

necessities amenities

Local media -0.000 -0.002*** 0.009*** 0.007*** 0.007***(0.000) (0.000) (0.001) (0.001) (0.001)

Observations 42,595 42,595 42,595 42,595 42,595

Notes: All specifications include year fixed effects, and are estimated using OLS. Standard errors areclustered by municipality. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

3

Figure 1: Sample ASF audit report (page 1)

Notes: Extracted from the ASF audit report on the use of FIMS funds by the municipal government ofthe municipality of Ajalpan in the state of Puebla in 2008. The red squares indicate the lines where theASF reports the FISM funds spent “in an unauthorized manner” and the share spent on projects “notbenefiting the poor.”

4

Figure 2: Sample ASF audit report (page 2)

Note: See Figure 1.

5

4 Additional results

Table 2 shows that controlling for the the total number of non-local media stations does not affectbalance in our GRD design. This is not surprising since the GRD breaks the correlation betweenthe number of local media stations and non-local media stations. Nevertheless, this represents animportant placebo further demonstrating that we not picking up differences in non-local mediastations. Furthermore, the balance tests suggest that estimates for non-local media are unlikely tobe biased too.

Table 3 presents our DD estimates for the GRD sample. Due to the significant decline in samplesize, and the fact that the GRD estimates only identify off within-match variation (and thus onlyfrom matches cross municipality borders), our estimates are considerably noisier. Nevertheless,the point estimates are very similar to those in the Table 4 of the main paper. The similarity is notsurprising given that the full and GRD samples are similar in composition (see Tables 1 and 3).

6

Tabl

e2:

GR

Dba

lanc

ech

ecks

,con

trol

ling

forn

on-l

ocal

med

iash

are

Una

utho

rize

dC

orru

ptC

orru

ptSp

endi

ngno

tN

otN

otR

egis

tere

dTu

rnou

tsp

endi

ngQ

3Q

4on

poor

poor

Q3

poor

Q4

vote

rs(l

ag)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Loc

alm

edia

-0.0

010.

001

-0.0

03-0

.001

0.00

9-0

.001

15.8

26-0

.001

(0.0

03)

(0.0

03)

(0.0

07)

(0.0

04)

(0.0

07)

(0.0

05)

(14.

406)

(0.0

01)

Non

-loc

alm

edia

0.00

0-0

.002

0.00

2-0

.000

-0.0

02-0

.004

-18.

155

0.00

0(0

.002

)(0

.002

)(0

.005

)(0

.003

)(0

.004

)(0

.004

)(1

3.98

8)(0

.001

)

PAN

PRI

PRD

Shar

eSh

are

Shar

ew

ithSh

are

with

Shar

ew

ithin

cum

bent

incu

mbe

ntin

cum

bent

empl

oyed

illite

rate

hous

ehol

dho

useh

old

inte

rnet

nece

ssiti

esam

eniti

es(9

)(1

0)(1

1)(1

2)(1

3)(1

4)(1

5)(1

6)

Loc

alm

edia

0.02

3***

-0.0

08-0

.015

***

0.00

0-0

.000

-0.0

000.

001

0.00

2**

(0.0

07)

(0.0

06)

(0.0

06)

(0.0

00)

(0.0

00)

(0.0

01)

(0.0

01)

(0.0

01)

Non

-loc

alm

edia

0.00

7-0

.003

-0.0

040.

000*

0.00

0-0

.002

***

0.00

00.

001

(0.0

05)

(0.0

04)

(0.0

04)

(0.0

00)

(0.0

00)

(0.0

01)

(0.0

00)

(0.0

01)

Not

es:A

llsp

ecifi

catio

nsin

clud

em

atch

and

year

fixed

effe

cts,

use

upto

two

poss

ible

mat

ches

,and

are

estim

ated

usin

gO

LS.

All

regr

essi

ons

have

19,2

30ob

serv

atio

ns.

Stan

dard

erro

rsar

ecl

uste

red

bym

unic

ipal

ity-y

ear.

*de

note

sp<

0.1,

**de

note

sp<

0.05

,**

*de

note

sp<

0.01

.

7

Table 3: The effect of audits revealing corruption (GRD sample)

Change in incumbent vote share Incumbent re-elected(1) (2) (3) (4) (5) (6)

Audit 0.015 0.007 0.027 0.121 0.078 0.174(0.026) (0.030) (0.033) (0.096) (0.094) (0.111)

Corrupt 0.068 0.076* 0.509*** 0.499***(0.044) (0.041) (0.166) (0.166)

Audit × Corrupt -0.054 -0.060 -0.240 -0.206(0.051) (0.049) (0.250) (0.256)

Not poor 0.018 0.038 -0.016 0.135(0.047) (0.051) (0.135) (0.156)

Audit × Not poor -0.034 -0.038 -0.231 -0.311(0.068) (0.070) (0.283) (0.274)

Observations 19,230 19,230 19,230 19,230 19,230 19,230R2 0.51 0.50 0.51 0.73 0.70 0.73

Notes: All specifications include pre-treatment controls and match and year fixed effects, use up to twopossible matches, and are estimated using OLS. All regressions have 19,230 observations except PanelA which contains 16,947 observations. Standard errors are clustered by municipality-year. * denotesp < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

8